The paper presents a framework for mitigating unwanted biases in machine learning models by incorporating an adversarial learning approach. The method involves training a predictor to predict an output variable \( Y \) from an input variable \( X \) while simultaneously training an adversary to predict a protected variable \( Z \). The objective is to maximize the predictor's ability to predict \( Y \) while minimizing the adversary's ability to predict \( Z \). This approach is flexible and can be applied to various definitions of fairness, including demographic parity, equality of odds, and equality of opportunity, as well as to a wide range of gradient-based learning models, including regression and classification tasks. The paper defines these fairness measures and explains how the adversarial technique can achieve them. For demographic parity, the adversary tries to predict \( Z \) from \( \hat{Y} \), and the predictor's weights are updated to reduce the information about \( Z \) in \( \hat{Y} \). For equality of odds, the adversary is given both \( \hat{Y \) and the true label \( Y \), limiting the information about \( Z \) in \( \hat{Y} \). The method is demonstrated on two scenarios: a toy scenario and the UCI Adult Dataset. In the toy scenario, a logistic regression model is trained to predict \( y \) while being unbiased with respect to the protected variable \( r \). In the UCI Adult Dataset, the model predicts income brackets while enforcing equality of odds. The results show that the debiased models achieve high accuracy while maintaining fairness. The paper also discusses theoretical guarantees and practical challenges, such as convergence issues, and proposes a simple adversary model that can be used regardless of the predictor's complexity. The method is shown to be effective in both supervised learning tasks and debiasing word embeddings, demonstrating its versatility and potential for real-world applications.The paper presents a framework for mitigating unwanted biases in machine learning models by incorporating an adversarial learning approach. The method involves training a predictor to predict an output variable \( Y \) from an input variable \( X \) while simultaneously training an adversary to predict a protected variable \( Z \). The objective is to maximize the predictor's ability to predict \( Y \) while minimizing the adversary's ability to predict \( Z \). This approach is flexible and can be applied to various definitions of fairness, including demographic parity, equality of odds, and equality of opportunity, as well as to a wide range of gradient-based learning models, including regression and classification tasks. The paper defines these fairness measures and explains how the adversarial technique can achieve them. For demographic parity, the adversary tries to predict \( Z \) from \( \hat{Y} \), and the predictor's weights are updated to reduce the information about \( Z \) in \( \hat{Y} \). For equality of odds, the adversary is given both \( \hat{Y \) and the true label \( Y \), limiting the information about \( Z \) in \( \hat{Y} \). The method is demonstrated on two scenarios: a toy scenario and the UCI Adult Dataset. In the toy scenario, a logistic regression model is trained to predict \( y \) while being unbiased with respect to the protected variable \( r \). In the UCI Adult Dataset, the model predicts income brackets while enforcing equality of odds. The results show that the debiased models achieve high accuracy while maintaining fairness. The paper also discusses theoretical guarantees and practical challenges, such as convergence issues, and proposes a simple adversary model that can be used regardless of the predictor's complexity. The method is shown to be effective in both supervised learning tasks and debiasing word embeddings, demonstrating its versatility and potential for real-world applications.